[test-suite] Add Image Processing Kernels Using Benchmark Library: Dither Algorithms

[proton]

Summary

Apply D49339 befor this patch. as that contains some utilities functions required by this.

[homerdin]

MicroBenchmarks/ImageProcessing/Dither/floydDitherKernel.cpp

9 ↗

(On Diff #156126)

WhiteSpace

10–11 ↗

(On Diff #156126)

Commented Code

MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.cpp

17 ↗

(On Diff #156126)

Commented Code

52 ↗

(On Diff #156126)

You could remove the M==2, M==3 and M==8 conditions since M is defined as 4? That or you could pass another argument to set M and have a test for each size image with each value of M.

[Meinersbur]

MicroBenchmarks/ImageProcessing/Dither/floydDitherKernel.cpp

9 ↗

(On Diff #157593)

This file does not use C++ features, so you can use C99 variable length array parameters.

43 ↗

(On Diff #157593)

temp1 is only used in the else case. No need to read the outputImage in all cases.

[proton]

Changes: Used C99 VLAs in kernels.

[Meinersbur]

MicroBenchmarks/ImageProcessing/Dither/main.cpp

37

Does this work? It dereferences the pointer returned by malloc (which should be the uninitialized first bytes of the memory allocated be malloc) and then casts it to a pointer. Id' expect this to result in a segfault.

MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.c

14

Do you need m to by a variable? It is only passed the constant 4 in this benchmark. The code would be more optimizable if the compiler knew the constant.

Alternatively, you could run the kernel several times with different values of m.

[proton]

MicroBenchmarks/ImageProcessing/Dither/main.cpp

37

Yes, It works properly because the pointer returned by malloc is casted as int (*) [][]. So to access the array we have to first dereference the pointer (use it as (*inputImage)[i][j]).

[proton]

Changes: update m along with image size when called with the benchmark library. For verification, only m=4 is used.

[proton]

changed array type from int (*) [][] to int * in main.cpp

[Meinersbur]

What is the execution time of this?

MicroBenchmarks/ImageProcessing/Dither/main.cpp

37

Looks much better now

100

Could you add a comment that this exists to avoid the computation being optimized away?

149

Here as well

[proton]

added some comments

[Meinersbur]

Did you check the execution time?

[MatzeB]

test-suite/trunk/MicroBenchmarks/ImageProcessing/Dither/main.cpp

144–154 ↗

(On Diff #159273)

I just noticed an odd effect here. A benchmark run now gives us 12 results for the same function running.

test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/512/3
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/512/2
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/256/3
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/256/2
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/128/2
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/128/3
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/256/4
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/512/4
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/128/4
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/512/8
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/128/8
test-suite :: MicroBenchmarks/ImageProcessing/Dither/Dither.test:BENCHMARK_ORDERED_DITHER/256/8

I just had a change that was mostly improving things but happened to slightly regress this benchmark. Unfortunately this effect is now multiplied by 12 so this benchmark has a far bigger weight than others...

[Meinersbur]

test-suite/trunk/MicroBenchmarks/ImageProcessing/Dither/main.cpp

144–154 ↗

(On Diff #159273)

Do you suggest to reduce the number of variants?

[proton]

test-suite/trunk/MicroBenchmarks/ImageProcessing/Dither/main.cpp

144–154 ↗

(On Diff #159273)

We can reduce the number of runs here but I think instead of hardcoding it here (like it is now), we can pass it as an argument from cmake file.
Earlier I dropped this idea because I remember someone mentioning that multiple test size is not used in the test suite now so it seemed unnecessary then.
This can help someone who may want to see the effect of some optimization (say tiling) on different input Matrix sizes and different zoom ratio here.
For hardcoding it, we can remove the multiple zoom ratio here
from

b->Args({i, 2});
b->Args({i, 3});
b->Args({i, 4});
b->Args({i, 8});

to

b->Args({i, 2});
b->Args({i, 8});

See Line: 107

MicroBenchmarks/ImageProcessing/CMakeLists.txt

+add_subdirectory(Dither)

MicroBenchmarks/ImageProcessing/Dither/CMakeLists.txt

+set(IMAGEPROC_UTILS MicroBenchmarks/ImageProcessing/utils)
+list(APPEND CPPFLAGS -I ${CMAKE_SOURCE_DIR}/${IMAGEPROC_UTILS} -std=c++11)
+
+llvm_test_verify("${CMAKE_SOURCE_DIR}/HashProgramOutput.sh ${CMAKE_CURRENT_BINARY_DIR}/orderedOutput.txt")
+llvm_test_verify("${FPCMP} ${CMAKE_CURRENT_BINARY_DIR}/orderedOutput.txt ${CMAKE_CURRENT_SOURCE_DIR}/orderedDither.reference_output")
+
+llvm_test_verify("${CMAKE_SOURCE_DIR}/HashProgramOutput.sh ${CMAKE_CURRENT_BINARY_DIR}/floydOutput.txt")
+llvm_test_verify("${FPCMP} ${CMAKE_CURRENT_BINARY_DIR}/floydOutput.txt ${CMAKE_CURRENT_SOURCE_DIR}/floydDither.reference_output")
+
+llvm_test_run(WORKDIR ${CMAKE_CURRENT_BINARY_DIR})
+llvm_test_executable(Dither main.cpp orderedDitherKernel.c  floydDitherKernel.c ../utils/ImageHelper.cpp ../utils/glibc_compat_rand.c)
+
+target_link_libraries(Dither benchmark)

MicroBenchmarks/ImageProcessing/Dither/dither.h

+/**
+  Pankaj Kukreja
+  github.com/proton0001
+  Indian Institute of Technology Hyderabad
+*/
+#ifndef _DITHER_H_
+#define _DITHER_H_
+
+#define MaxGray 255
+#define MXGRAY 256
+
+#define HEIGHT 512
+#define WIDTH 512
+
+#endif /* _DITHER_H_ */

MicroBenchmarks/ImageProcessing/Dither/floydDither.reference_output

+23473c2d34c91e33eaa3d5008ce3640e

MicroBenchmarks/ImageProcessing/Dither/floydDitherKernel.c

+/**
+  Source: https://imagej.net/Dithering
+  Modified by Pankaj Kukreja (github.com/proton0001)
+  Indian Institute of Technology Hyderabad
+*/
+#include "dither.h"
+void floydDitherKernel(int height, int width, int inputImage[HEIGHT][WIDTH],
+                       int outputImage[height][width]) {
+  for (int i = 0; i < height; i++) {
+    for (int j = 0; j < width; j++) {
+      outputImage[i][j] = inputImage[i][j];
+    }
+  }
+
+  int err;
+  int a, b, c, d;
+
+  for (int i = 1; i < height - 1; i++) {
+    for (int j = 1; j < width - 1; j++) {
+      if (outputImage[i][j] > 127) {
+        err = outputImage[i][j] - 255;
+        outputImage[i][j] = 255;
+      } else {
+        err = outputImage[i][j] - 0;
+        outputImage[i][j] = 0;
+      }
+      a = (err * 7) / 16;
+      b = (err * 1) / 16;
+      c = (err * 5) / 16;
+      d = (err * 3) / 16;
+
+      int temp1 = (outputImage[i][j + 1] + a);
+      if (temp1 > 255) {
+        outputImage[i][j + 1] = 255;
+      } else if (temp1 < 0) {
+        outputImage[i][j + 1] = 0;
+      } else {
+        outputImage[i][j + 1] = temp1;
+      }
+
+      int temp2 = (outputImage[i + 1][j + 1] + b);
+      if (temp2 > 255) {
+        outputImage[i + 1][j + 1] = 255;
+      } else if (temp2 < 0) {
+        outputImage[i + 1][j + 1] = 0;
+      } else {
+        outputImage[i + 1][j + 1] = temp2;
+      }
+
+      int temp3 = outputImage[i + 1][j + 0] + c;
+      if (temp3 > 255) {
+        outputImage[i + 1][j + 0] = 255;
+      } else if (temp3 < 0) {
+        outputImage[i + 1][j + 0] = 0;
+      } else {
+        outputImage[i + 1][j + 0] = temp3;
+      }
+
+      int temp4 = outputImage[i + 1][j - 1] + d;
+      if (temp4 > 255) {
+        outputImage[i + 1][j - 1] = 255;
+      } else if (temp4 < 0) {
+        outputImage[i + 1][j - 1] = 0;
+      } else {
+        outputImage[i + 1][j - 1] = temp4;
+      }
+    }
+  }
+}

MicroBenchmarks/ImageProcessing/Dither/main.cpp

+/**
+  Pankaj Kukreja
+  github.com/proton0001
+  Indian Institute of Technology Hyderabad
+*/
+#include "ImageHelper.h"
+#include "dither.h"
+#include <cmath>
+#include <iostream> // std::cerr
+
+#define BENCHMARK_LIB
+#ifdef BENCHMARK_LIB
+#include "benchmark/benchmark.h"
+#endif
+
+int *inputImage;
+extern "C" {
+void orderedDitherKernel(int height, int width, int *inpImage, int *outImage,
+                         int *temp, int n, int m);
+void floydDitherKernel(int height, int width, int *inpImage, int *outImage);
+}
+int main(int argc, char **argv) {
+#ifdef BENCHMARK_LIB
+  ::benchmark::Initialize(&argc, argv);
+#endif
+
+  const char *orderedOutputFilename = (char *)"./orderedOutput.txt";
+  const char *floydOutputFilename = (char *)"./floydOutput.txt";
+  inputImage = (int *)malloc(sizeof(int) * HEIGHT * WIDTH);
+  if (inputImage == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(1);
+  }
+  initializeRandomImage(inputImage, HEIGHT, WIDTH);
+
+#ifdef BENCHMARK_LIB
+  ::benchmark::RunSpecifiedBenchmarks();
+#endif
+  int *outputImage = (int *)malloc(sizeof(int) * HEIGHT * WIDTH);
+  int *temp = (int *)malloc(sizeof(int) * HEIGHT * WIDTH);
+  if (outputImage == NULL || temp == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(1);
+  }
+  orderedDitherKernel(HEIGHT, WIDTH, inputImage, outputImage, temp, 16, 4);
+  saveImage(outputImage, orderedOutputFilename, HEIGHT, WIDTH);
+  floydDitherKernel(HEIGHT, WIDTH, inputImage, outputImage);
+
+  for (int i = 0; i < HEIGHT; i++) {
+    outputImage[(i)*WIDTH + 0] = 0;
+    outputImage[(i)*WIDTH + WIDTH - 1] = 0;
+  }
+
+  for (int j = 0; j < WIDTH; j++) {
+    outputImage[(0) * WIDTH + j] = 0;
+    outputImage[(HEIGHT - 1) * WIDTH + j] = 0;
+  }
+
+  saveImage(outputImage, floydOutputFilename, HEIGHT, WIDTH);
+  free(temp);
+  free(outputImage);
+  free(inputImage);
+  return EXIT_SUCCESS;
+}
+
+#ifdef BENCHMARK_LIB
+void BENCHMARK_ORDERED_DITHER(benchmark::State &state) {
+  int height = state.range(0);
+  int width = state.range(0);
+  int m = state.range(1);
+  int n = pow(m, 2);
+  int *outputImage = (int *)malloc(sizeof(int) * height * width);
+  int *temp = (int *)malloc(sizeof(int) * height * width);
+
+  if (outputImage == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(1);
+  }
+  /* This call is to warm up the cache */
+  orderedDitherKernel(height, width, inputImage, outputImage, temp, n, m);
+
+  for (auto _ : state) {
+    orderedDitherKernel(height, width, inputImage, outputImage, temp, n, m);
+  }
+  /* Since we are not passing state.range as 20 this if case will always be
+   * false. This call is to make compiler think that outputImage may be used
+   * later so that above kernel calls will not optimize out */
+  if (state.range(0) == 20) {
+    saveImage(outputImage, (const char *)"failedCase.txt", height, width);
+  }
+  free(temp);
+  free(outputImage);
+}
+
+#if (HEIGHT < WIDTH)
+#define MINIMUM_DIM HEIGHT
+#else
+#define MINIMUM_DIM WIDTH
+#endif
+
+static void CustomArguments(benchmark::internal::Benchmark *b) {
+  int limit = MINIMUM_DIM;
+  int start = 1;
+  if (limit > 128) {
+    start = 128;
+  }
+  for (int i = start; i <= limit; i <<= 1) {
+    b->Args({i, 2});
+    b->Args({i, 3});
+    b->Args({i, 4});
+    b->Args({i, 8});
+  }
+}
+BENCHMARK(BENCHMARK_ORDERED_DITHER)
+    ->Apply(CustomArguments)
+    ->Unit(benchmark::kMicrosecond);
+
+void BENCHMARK_FLOYD_DITHER(benchmark::State &state) {
+
+  int height = state.range(0);
+  int width = state.range(0);
+
+  int *outputImage = (int *)malloc(sizeof(int) * height * width);
+
+  if (outputImage == NULL) {
+    std::cerr << "Insufficient memory\n";
+    exit(1);
+  }
+  /* This call is to warm up the cache */
+  floydDitherKernel(height, width, inputImage, outputImage);
+  for (auto _ : state) {
+    floydDitherKernel(height, width, inputImage, outputImage);
+  }
+  /* Since we are not passing state.range as 20 this if case will always be
+   * false. This call is to make compiler think that outputImage may be used
+   * later so that above kernel calls will not optimize out */
+  if (state.range(0) == 20) {
+    saveImage(outputImage, (const char *)"failedCase.txt", height, width);
+  }
+
+  free(outputImage);
+}
+
+#if MINIMUM_DIM > 128
+BENCHMARK(BENCHMARK_FLOYD_DITHER)
+    ->RangeMultiplier(2)
+    ->Range(128, MINIMUM_DIM)
+    ->Unit(benchmark::kMicrosecond);
+#else
+BENCHMARK(BENCHMARK_FLOYD_DITHER)
+    ->RangeMultiplier(2)
+    ->Range(1, MINIMUM_DIM)
+    ->Unit(benchmark::kMicrosecond);
+#endif
+
+#endif

MicroBenchmarks/ImageProcessing/Dither/orderedDither.reference_output

+7b339ccc04bbaebf30f44f4f3129756f

MicroBenchmarks/ImageProcessing/Dither/orderedDitherKernel.c

+/**
+  Source: github -> https://github.com/brianwu02/ImageProcessing.git
+  Modified by Pankaj Kukreja (github.com/proton0001)
+  Indian Institute of Technology Hyderabad
+*/
+#include "dither.h"
+#include <math.h> // pow
+
+#define GAMMA 0.5
+
+void orderedDitherKernel(int height, int width, int inputImage[HEIGHT][WIDTH],
+                         int outputImage[height][width],
+                         int temp[height][width], int n, int m) {
+  int scale;
+
+  for (int i = 0; i < height; i++) {
+    for (int j = 0; j < width; j++) {
+      temp[i][j] =
+          (int)(pow((double)inputImage[i][j] / 255.0, (1.0 / GAMMA)) * 255.0);
+    }
+  }
+
+  scale = 256 / n;
+  for (int i = 0; i < height; i++) {
+    for (int j = 0; j < width; j++) {
+      outputImage[i][j] = (int)(scale * (temp[i][j] / scale)) / scale;
+    }
+  }
+
+  if (m == 2) {
+    int dither[2][2] = {{0, 2}, {3, 1}};
+    for (int y = 0; y < height; y++) {
+      for (int x = 0; x < width; x++) {
+        int i = x % m;
+        int j = y % m;
+        outputImage[y][x] = ((outputImage[y][x] > dither[i][j]) ? 255 : 0);
+      }
+    }
+  } else if (m == 3) {
+    int dither[3][3] = {{6, 8, 4}, {1, 0, 3}, {5, 2, 7}};
+    for (int y = 0; y < height; y++) {
+      for (int x = 0; x < width; x++) {
+        int i = x % m;
+        int j = y % m;
+        outputImage[y][x] = ((outputImage[y][x] > dither[i][j]) ? 255 : 0);
+      }
+    }
+  } else if (m == 4) {
+    int dither[4][4] = {
+        {0, 8, 2, 10}, {12, 4, 14, 6}, {3, 11, 1, 9}, {15, 7, 13, 5}};
+    for (int y = 0; y < height; y++) {
+      for (int x = 0; x < width; x++) {
+        int i = x % m;
+        int j = y % m;
+        outputImage[y][x] = ((outputImage[y][x] > dither[i][j]) ? 255 : 0);
+      }
+    }
+  } else if (m == 8) {
+    int dither[8][8] = {
+        {0, 48, 12, 60, 3, 51, 15, 63}, {32, 16, 44, 28, 35, 19, 47, 31},
+        {8, 56, 4, 52, 11, 59, 7, 55},  {40, 24, 36, 20, 43, 27, 39, 23},
+        {2, 50, 14, 62, 1, 49, 13, 61}, {34, 18, 46, 30, 33, 17, 45, 29},
+        {10, 58, 6, 54, 9, 57, 5, 53},  {42, 26, 38, 22, 41, 25, 37, 21}};
+    for (int y = 0; y < height; y++) {
+      for (int x = 0; x < width; x++) {
+        int i = x % m;
+        int j = y % m;
+        outputImage[y][x] = ((outputImage[y][x] > dither[i][j]) ? 255 : 0);
+      }
+    }
+  }
+}